MyAgentRoam

Your agents.
Wherever
you roam.

A self-hosted Agent workbench.
Let Agents work on your machines. Pick up from a browser, wherever you are.

Start deploying View source
Supported Agents
  • MAR Agent
  • Codex
  • Claude Code
  • OpenCode
View the full desktop screenshot View the full mobile screenshot
Desktop and mobile interface views

Follow tasks on your phone.

Projects, dependencies and execution stay on your own Windows, Linux or macOS machines. Switch registered Nodes and workspaces from one workbench, without setting up each browsing device.

  1. Start at your deskChoose a workspace. Give your Agent a task.
  2. Check in by phoneFollow progress, receive notifications and steer.
  3. Back at your deskReview the changes. Keep the same task moving.

Enable Computer Use for MAR Agent when an engineering task needs a desktop GUI rather than a command line.

Computer Use requires an available interactive desktop and explicit enablement for the current run. Your self-hosted Server must be reachable and the relevant Node online for live work. Complex diffs and extended terminal work suit a desktop better. The author has no Mac; macOS has not yet been tested.

MyAgentRoam mobile browser workbench
The mobile browser workbench

Projects and execution stay on your machines.

Projects, dependencies, Git and toolchains stay on the Node. Your browser accesses authorized workspaces through the Server to assign tasks, check progress and review changes. Execution stays on the Node.

  1. BrowserObserve & steer
  2. ServerIdentity & access
  3. NodeReal workspace

Coordinate work across Agents.

Use marctl to coordinate independent MAR Agent, Codex, Claude Code or OpenCode sessions on the same Node and workspace. Divide research, implementation and review, exchange messages and share artifacts while retaining separate contexts.

Unlike a Subagent within a parent execution, these are independent sessions. Operations follow the capabilities of each Runner; this is not an automatic workflow engine.

Agent Network · Example roles
  1. Research
  2. Implement
  3. Review
MyAgentRoam workbench showing Agent collaboration

Manage context through long tasks.

Long tasks need more than a one-off compaction. MAR Agent combines incremental reclamation, phase-aware compaction and capacity protection, with visible context state and manual control.

Reclaim historical context incrementally
Context GC incrementally reclaims eligible history, reducing old details carried into subsequent requests.
Compact at a useful transition
Smart Compaction lets the Agent request background compaction at suitable phase transitions. Fixed capacity protection works independently of that request.
Inspect usage and compact manually
Inspect reported context usage, compaction status and plans; compact manually when needed. Compaction does not delete conversation history or promise unlimited, lossless memory.
  • Context GC
  • Steer
  • Code Mode
  • Subagents
  • Agent Network
  • Auto Access
  • Computer Use
Explore MAR Agent capabilities and requirements